Abstract
Hydrogeological structures often contain heterogeneous patterns (i.e., multiple-type facies and aquifers), making the characterization of multiple type structures a challenging task in geological modeling. The existing deep learning-based methods are insufficient to learn multiple-pattern features within a single training process, which affects the accuracy and efficiency of aquifer structure characterization. In this work, a Diffusion-based Multi-Category Generation (DMCG) network is proposed to unify the learning of multiple patterns contained in different types of aquifer structures. This method integrates geological observations (e.g., drilling and cross-section) and category labels and uses U-Net to establish the relationship between input data and spatial features. The embedding of category labels is the key factor for DMCG to learn different type aquifers. Moreover, a vector quantized variational autoencoder is employed to implement a latent diffusion model, which can reduce the memory requirement for training and improve characterization efficiency. Two datasets containing various patterns of aquifer structures are used to validate the performance of DMCG. Variograms, multiple dimensional scaling maps, cumulative proportional distribution function, and other metrics are used to visualize the differences between realizations and references in terms of spatial heterogeneity and attribute proportion. Experimental results demonstrate that DMCG can characterize multiple-type aquifer structures under the constraints of both conditioning data and category labels. The comparative test illustrates that DMCG achieves better performance to a multiple-point geostatistical algorithm (Quick Sampling, QS) and recent diffusion-based method in reproducing multiple-type hydrogeological structures.